You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
It looks like adding OpenCL support to this wouldn't take much effort, though I'm still so early at this that I'm not comfortable offering a PR yet; have you any familiarity with torchcl and clnn?
Advantage: newcomers (who seem to be the target audience) can use CPU integrated graphics to train, and can scale with much, much cheaper GPUs with similar performance. The CPU Graphics angle makes it more suitable for workshop/teaching use, where people will bring laptops and are unlikely to have nVidia GPUs.
The text was updated successfully, but these errors were encountered:
The difference, as far as I know so far, is simply that for OpenCL you import torchcl and clnn, and then wherever you'd call :cuda() you instead call :cl(). Really it'd be cool to up-stream something to torch itself, so that you could just call :gpu() and have it try loading either way..
Very neat and tidy code!
It looks like adding OpenCL support to this wouldn't take much effort, though I'm still so early at this that I'm not comfortable offering a PR yet; have you any familiarity with
torchcl
andclnn
?Advantage: newcomers (who seem to be the target audience) can use CPU integrated graphics to train, and can scale with much, much cheaper GPUs with similar performance. The CPU Graphics angle makes it more suitable for workshop/teaching use, where people will bring laptops and are unlikely to have nVidia GPUs.
The text was updated successfully, but these errors were encountered: